vix.ing · top · new · best · stats

Generative Adversarial Networks

2014/06/10 by Ian Goodfellow, Ian J. Goodfellow, Goodfellow, Ian J. +14 · 7 voices · 4,605 citations
Computer Science · Mathematics · #Artificial intelligence #Artificial neural network #Computer science #Discriminative model #Generative Adversarial Networks and Image Synthesis #Generative grammar #Generative model #Image Processing and 3D Reconstruction #Inference #Machine learning #Mathematical optimization #Mathematics #Minimax #Mistake #Music and Audio Processing #Perceptron #Sample (material) #cs.LG #stat.ML

paper · pdf · doi:10.48550/arxiv.1406.2661

published in arXiv (Cornell University) (Cornell University)

arxiv created 2014/06/10 · openalex publication_date 2014/06/10 · arxiv updated 2014/06/11 · openalex created_date 2022/10/01 · openalex updated_date 2026/08/06

Abstract

Large Language Models (LLMS) rely on Key-Value (KV) caches to store attention context during autoregressive decoding. In long-sequence settings, the KV cache can consume large amounts of VRAM and become a practical bottleneck for throughput . We introduce KVHALO, an auxiliary reconstruction model that restores higher-fidelity KV tensors from a compressed cache state when required, reducing persistent memory footprint during inference. In our evaluation, KVHALO achieves up to 91.85% directional cosine alignment at convergence and reduces long-context degradation relative to a low-bit baseline under our stress-test workloads. We used HRM instead of other architectures, which allowed for higher-quality results in only 18,600 steps.

Cited by

Discussions

Related